is it possible to finetune from original weights?

[yyfcc17]

or we need to train from scratch every time?

[Hexuanfang]

hi, did you figure it out? i have the some question

[yyfcc17]

Yes, i tried to reuse the original weights of a conv net trained on MNIST dataset. when there are just two layers, and alpha set to 0.5, use the original weights to init the octave conv parameters, can recover an accuracy of 91% vs 98% of the original.
but when it comes to 3 layers of convolution, the accuracy drops to 17%.
i think it will get much worse after the network goes deeper.
so my answer is no. currently, it seems that we cannot adapt the finetune strategy to use OctaveConv.

[Hexuanfang]

thanks a million

…

---Original--- From: "yafeng"<[email protected]&gt; Date: Fri, May 15, 2020 19:19 PM To: "facebookresearch/OctConv"<[email protected]&gt;; Cc: "Comment"<[email protected]&gt;;"Hexuanfang"<[email protected]&gt;; Subject: Re: [facebookresearch/OctConv] is it possible to finetune from original weights? (#14) Yes, i tried to reuse the original weights of a conv net trained on MNIST dataset. when there are just two layers, and alpha set to 0.5, use the original weights to init the octave conv parameters, can recover an accuracy of 91% vs 98% of the original. but when it comes to 3 layers of convolution, the accuracy drops to 17%. i think it will get much worse after the network goes deeper. so my answer is no. currently, it seems that we cannot adapt the finetune strategy to use OctaveConv. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.